DreamControl: Control-Based Text-to-3D Generation with 3D Self-Prior

TL;DR

This work proposes a two-stage 2D-lifting framework, namely DreamControl, which optimizes coarse NeRF scenes as 3D self-prior and then generates fine-grained objects with control-based score distillation using conditional LoRA and weighted score to optimize detailed textures.

Abstract

3D generation has raised great attention in recent years. With the success of text-to-image diffusion models, the 2D-lifting technique becomes a promising route to controllable 3D generation. However, these methods tend to present inconsistent geometry, which is also known as the Janus problem. We observe that the problem is caused mainly by two aspects, i.e., viewpoint bias in 2D diffusion models and overfitting of the optimization objective. To address it, we propose a two-stage 2D-lifting framework, namely DreamControl, which optimizes coarse NeRF scenes as 3D self-prior and then generates fine-grained objects with control-based score distillation. Specifically, adaptive viewpoint sampling and boundary integrity metric are proposed to ensure the consistency of generated priors. The priors are then regarded as input conditions to maintain reasonable geometries, in which conditional LoRA and weighted score are further proposed to optimize detailed textures. DreamControl can generate high-quality 3D content in terms of both geometry consistency and texture fidelity. Moreover, our control-based optimization guidance is applicable to more downstream tasks, including user-guided generation and 3D animation. The project page is available at https://github.com/tyhuang0428/DreamControl.

Figures (14)

• Figure 1: Main causes of inconsistent 3D generation. Images sampled by 2D diffusion models are biased in viewpoint distribution. The generation confidence decreases as the viewpoint turns from front to back. 3D representations are thus gradually overfitted to the highest probability image during the optimization, generating artifacts as shown in red b-boxes.
• Figure 2: DreamControl can generate diverse 3D content with high-consistency geometries and high-fidelity textures. Beyond text-to-3D generation, our control-based guidance is applicable to controllable generation tasks, including user-guided generation and 3D animation.
• Figure 3: Overview of DreamControl. In the first stage, a coarse NeRF is optimized as a 3D self-prior $\hat{\theta}$, in which an adaptive viewpoint sampling $p^{\ast}$ and a boundary integrity metric $\Delta_{\mathbf{r}}$ are proposed to alleviate inconsistent generation. The prior $\hat{\theta}$ is then sent to the second stage as an input edge condition $\hat{x}_t$, in which a control-based score distillation can generate fine-grained textures and maintain geometries in the prior. A Conditional LoRA and a weighted score are further proposed to stabilize the optimization process.
• Figure 4: Qualitative results. Compared with other methods, DreamControl enjoys high-consistency geometry and high-fidelity texture.
• Figure 5: User-guided generation. DreamControl is flexible to loose input conditions, generating fine-grained content with a 3D sketch or even a coarse layout.